Merge pull request #318 from opencompl/sasha/plane-hack

add fill baseline, linalg, and linalg_xdsl f32
opencompl · Oct 22, 2024 · 7a3c617 · 7a3c617
2 parents 1fc89dd + db5f48a
commit 7a3c617
Show file tree

Hide file tree

Showing 14 changed files with 36 additions and 9 deletions.
diff --git a/Snakefile b/Snakefile
@@ -130,6 +130,9 @@ TESTSET_FAST = [
     *expand(
         "relu/4x8xf32/{variant}", variant=["baseline", "linalg", "snrt", "snitch_stream"]
     ),
+    *expand(
+        "fill/4x8xf32/{variant}", variant=["baseline", "linalg", "linalg_xdsl"]
+    ),
     *expand(
         "sum/4x8xf32/{variant}", variant=["baseline", "snrt", "linalg", "linalg_xdsl"]
     ),

diff --git a/kernels/fill/baseline.c.template b/kernels/fill/baseline.c.template
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 
-void fill(const double x, double* restrict y) {
+void fill(const DTYPE x, DTYPE* restrict y) {
     for (uint32_t i = 0; i < M; ++i) {
         for (uint32_t j = 0; j < N; ++j) {
             // row-major accesses

diff --git a/kernels/fill/data.h.template b/kernels/fill/data.h.template
@@ -1,9 +1,20 @@
 #pragma once
 
+#define PRECISION {{precision}}
+#if PRECISION == 16
+#define DTYPE __fp16
+#elif PRECISION == 32
+#define DTYPE float
+#elif PRECISION == 64
+#define DTYPE double
+#else
+#error Unsupported precision
+#endif
+
 #define M {{M}}
 #define N {{N}}
 
-extern const double X;
-extern const double Y[M * N];
+extern const DTYPE X;
+extern const DTYPE Y[M * N];
 
 #define TEST_COUNT (M * N)
diff --git a/kernels/fill/linalg.mlir.template b/kernels/fill/linalg.mlir.template
@@ -1,6 +1,6 @@
 
-func.func public @fill(%X: f64,
-                       %Y: memref<{{M}}x{{N}}xf64> {llvm.noalias}) -> () {
-  linalg.fill ins(%X : f64) outs(%Y : memref<{{M}}x{{N}}xf64>) -> ()
+func.func public @fill(%X: f{{precision}},
+                       %Y: memref<{{M}}x{{N}}xf{{precision}}> {llvm.noalias}) -> () {
+  linalg.fill ins(%X : f{{precision}}) outs(%Y : memref<{{M}}x{{N}}xf{{precision}}>) -> ()
   return
 }
diff --git a/kernels/fill/main.c b/kernels/fill/main.c
@@ -5,15 +5,15 @@
 #include <math.h>
 
 // Kernel provided via external definition
-void fill(double x, double *y);
+void fill(DTYPE x, DTYPE *y);
 
 int main() {
     // Allocate shared local memory
     // By avoiding allocators and bumping by a known offset a base pointer
     // (snrt_l1_next()) that is the same for all the cores in the cluster, we are
     // essentially providing the same memory regions to all the cores in this cluster.
-    double local_x = X;
-    double *local_y = (double *)snrt_l1_next();
+    DTYPE local_x = X;
+    DTYPE *local_y = (DTYPE *)snrt_l1_next();
 
     snrt_cluster_hw_barrier();
 

diff --git a/results/kernels.csv b/results/kernels.csv
@@ -16,6 +16,9 @@ dense,8x8xf64,snrt,2723,3538,3535,2.630311614730878,0.0,512,706,1857,0.259272860
 fill,4x4xf64,baseline,50,760,757,0.0,0.0,0,1,0,0.02,0.058823529411764705,17,0,0,0.34,0,16,1.0,1.0,1,0.0,17,0.7391304347826086,6,0,0,0.12,0,711,0.0,0.46,0.0
 fill,4x4xf64,linalg,50,760,757,0.0,0.0,0,1,0,0.02,0.058823529411764705,17,0,0,0.34,0,16,1.0,1.0,1,0.0,17,0.7391304347826086,6,0,0,0.12,0,711,0.0,0.46,0.0
 fill,4x4xf64,linalg_xdsl,64,766,763,0.9444444444444444,0.0,0,18,17,0.28125,0.9,20,0,0,0.3125,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.3,14,0,0,0.21875,0,703,0.0,0.53125,0.0
+fill,4x8xf32,baseline,74,798,795,0.0,0.0,0,1,0,0.013513513513513514,0.030303030303030304,33,0,0,0.44594594594594594,0,32,1.0,1.0,1,0.0,33,0.8461538461538461,6,0,0,0.08108108108108109,0,725,0.0,0.527027027027027,0.0
+fill,4x8xf32,linalg,74,798,795,0.0,0.0,0,1,0,0.013513513513513514,0.030303030303030304,33,0,0,0.44594594594594594,0,32,1.0,1.0,1,0.0,33,0.8461538461538461,6,0,0,0.08108108108108109,0,725,0.0,0.527027027027027,0.0
+fill,4x8xf32,linalg_xdsl,68,788,785,0.9444444444444444,0.0,0,18,17,0.2647058823529412,0.9,20,0,0,0.29411764705882354,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.3,14,0,0,0.20588235294117646,0,721,0.0,0.5,0.0
 matmul,1x20x5xf64,linalg_0_xdsl,4136,4896,4893,2.950980392156863,1.0,100,102,301,0.024661508704061894,0.20118343195266272,507,300,300,0.12258220502901354,0,105,1.0,1.0,1,0.0,507,0.19735305566368236,2062,0,0,0.4985493230174081,0,761,0.0,0.6211315280464217,0.0
 matmul,1x20x5xf64,linalg_1_xdsl,2050,2829,2826,2.8598130841121496,1.0,100,107,306,0.05219512195121951,0.3440514469453376,311,100,100,0.15170731707317073,0,100,1.0,1.0,1,0.0,311,0.22552574329224076,1068,0,0,0.5209756097560976,0,780,0.0,0.6726829268292683,0.0
 matmul,1x20x5xf64,linalg_2_xdsl,541,1320,1317,2.8598130841121496,1.0,100,107,306,0.1977818853974122,0.8842975206611571,121,5,5,0.22365988909426987,0,5,1.0,1.0,1,0.0,121,0.28205128205128205,308,0,0,0.5693160813308688,0,780,0.0,0.7929759704251387,0.0

diff --git a/results/kernels.fast.csv b/results/kernels.fast.csv
@@ -16,6 +16,9 @@ dense,8x8xf64,snrt,2723,3538,3535,2.630311614730878,0.0,512,706,1857,0.259272860
 fill,4x4xf64,baseline,50,760,757,0.0,0.0,0,1,0,0.02,0.058823529411764705,17,0,0,0.34,0,16,1.0,1.0,1,0.0,17,0.7391304347826086,6,0,0,0.12,0,711,0.0,0.46,0.0
 fill,4x4xf64,linalg,50,760,757,0.0,0.0,0,1,0,0.02,0.058823529411764705,17,0,0,0.34,0,16,1.0,1.0,1,0.0,17,0.7391304347826086,6,0,0,0.12,0,711,0.0,0.46,0.0
 fill,4x4xf64,linalg_xdsl,64,766,763,0.9444444444444444,0.0,0,18,17,0.28125,0.9,20,0,0,0.3125,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.3,14,0,0,0.21875,0,703,0.0,0.53125,0.0
+fill,4x8xf32,baseline,74,798,795,0.0,0.0,0,1,0,0.013513513513513514,0.030303030303030304,33,0,0,0.44594594594594594,0,32,1.0,1.0,1,0.0,33,0.8461538461538461,6,0,0,0.08108108108108109,0,725,0.0,0.527027027027027,0.0
+fill,4x8xf32,linalg,74,798,795,0.0,0.0,0,1,0,0.013513513513513514,0.030303030303030304,33,0,0,0.44594594594594594,0,32,1.0,1.0,1,0.0,33,0.8461538461538461,6,0,0,0.08108108108108109,0,725,0.0,0.527027027027027,0.0
+fill,4x8xf32,linalg_xdsl,68,788,785,0.9444444444444444,0.0,0,18,17,0.2647058823529412,0.9,20,0,0,0.29411764705882354,0,0,3.333333333333333,3.3333333333333335,1,0.0,6,0.3,14,0,0,0.20588235294117646,0,721,0.0,0.5,0.0
 matmul,1x20x5xf64,linalg_0_xdsl,4136,4896,4893,2.950980392156863,1.0,100,102,301,0.024661508704061894,0.20118343195266272,507,300,300,0.12258220502901354,0,105,1.0,1.0,1,0.0,507,0.19735305566368236,2062,0,0,0.4985493230174081,0,761,0.0,0.6211315280464217,0.0
 matmul,1x20x5xf64,linalg_1_xdsl,2050,2829,2826,2.8598130841121496,1.0,100,107,306,0.05219512195121951,0.3440514469453376,311,100,100,0.15170731707317073,0,100,1.0,1.0,1,0.0,311,0.22552574329224076,1068,0,0,0.5209756097560976,0,780,0.0,0.6726829268292683,0.0
 matmul,1x20x5xf64,linalg_2_xdsl,541,1320,1317,2.8598130841121496,1.0,100,107,306,0.1977818853974122,0.8842975206611571,121,5,5,0.22365988909426987,0,5,1.0,1.0,1,0.0,121,0.28205128205128205,308,0,0,0.5693160813308688,0,780,0.0,0.7929759704251387,0.0

diff --git a/results/pivoted.csv b/results/pivoted.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,667,858,308,,
 ddot 128xf64,956,965,,213,577
 dense 8x8xf64,3206,3530,,2741,2723
 fill 4x4xf64,50,50,64,,
+fill 4x8xf32,74,74,68,,
 matmul 4x16x8xf64,2495,2694,708,,
 matmul_transb 4x16x16xf32,3386,5038,,845,849
 matmul_transb 4x16x16xf64,,5142,1295,,

diff --git a/results/pivoted.fast.csv b/results/pivoted.fast.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,667,858,308,,
 ddot 128xf64,956,965,,213,577
 dense 8x8xf64,3206,3530,,2741,2723
 fill 4x4xf64,50,50,64,,
+fill 4x8xf32,74,74,68,,
 matmul 4x16x8xf64,2495,2694,708,,
 matmul_transb 4x16x16xf32,3386,5038,,845,849
 matmul_transb 4x16x16xf64,,5142,1295,,

diff --git a/results/pivoted_fpu.csv b/results/pivoted_fpu.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,0.22,0.17,0.58,,
 ddot 128xf64,0.13,0.13,,0.64,0.22
 dense 8x8xf64,0.20,0.18,,0.26,0.26
 fill 4x4xf64,0.02,0.02,0.28,,
+fill 4x8xf32,0.01,0.01,0.26,,
 matmul 4x16x8xf64,0.21,0.19,0.82,,
 matmul_transb 4x16x16xf32,0.21,0.20,,0.76,0.79
 matmul_transb 4x16x16xf64,,0.20,0.89,,

diff --git a/results/pivoted_fpu.fast.csv b/results/pivoted_fpu.fast.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,0.22,0.17,0.58,,
 ddot 128xf64,0.13,0.13,,0.64,0.22
 dense 8x8xf64,0.20,0.18,,0.26,0.26
 fill 4x4xf64,0.02,0.02,0.28,,
+fill 4x8xf32,0.01,0.01,0.26,,
 matmul 4x16x8xf64,0.21,0.19,0.82,,
 matmul_transb 4x16x16xf32,0.21,0.20,,0.76,0.79
 matmul_transb 4x16x16xf64,,0.20,0.89,,

diff --git a/results/pivoted_ipc.csv b/results/pivoted_ipc.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,0.40,0.45,0.80,,
 ddot 128xf64,0.95,0.94,,0.74,0.25
 dense 8x8xf64,0.51,0.55,,0.39,0.33
 fill 4x4xf64,0.46,0.46,0.53,,
+fill 4x8xf32,0.53,0.53,0.50,,
 matmul 4x16x8xf64,0.56,0.58,0.93,,
 matmul_transb 4x16x16xf32,0.95,0.55,,0.99,0.88
 matmul_transb 4x16x16xf64,,0.56,0.97,,

diff --git a/results/pivoted_ipc.fast.csv b/results/pivoted_ipc.fast.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3 4x4xf64,0.40,0.45,0.80,,
 ddot 128xf64,0.95,0.94,,0.74,0.25
 dense 8x8xf64,0.51,0.55,,0.39,0.33
 fill 4x4xf64,0.46,0.46,0.53,,
+fill 4x8xf32,0.53,0.53,0.50,,
 matmul 4x16x8xf64,0.56,0.58,0.93,,
 matmul_transb 4x16x16xf32,0.95,0.55,,0.99,0.88
 matmul_transb 4x16x16xf64,,0.56,0.97,,

diff --git a/results/regalloc.fast.csv b/results/regalloc.fast.csv
@@ -3,6 +3,7 @@ conv2d_d1_s1_3x3,4x4xf64,8,8
 ddot,128xf64,6,7
 dense,8x8xf64,5,11
 fill,4x4xf64,3,3
+fill,4x8xf32,3,3
 matmul,4x16x8xf64,8,8
 matmul_transb,4x16x16xf32,11,12
 matmul_transb,4x16x16xf64,8,8